gimmy256/afri-aya-vision

Afri-Aya Vision: Restructured for Multimodal & Adaption Fine-Tuning

This dataset is a restructured, multimodal Vision-Language (VLM) adaptation of [CohereLabsCommunity/afri-aya](https://huggingface.co/datasets/CohereLabsCommunity/afri-aya) (Giving Sight to African LLMs).

Why This Restructured Version?

The original Afri-Aya dataset stores multiple question-and-answer pairs per image inside a nested list column (qa_pairs). Fine-tuning platforms (such as Adaption, Unsloth, LLaVA, and standard VLM training harnesses) require:

1. 1.1 training row per question-answer pair (exploded rows).
2. 2.Direct image linkage via image_filename, relative image_path, and embedded PIL images.
3. 3.Clean prompt, completion, and native chat columns.

This dataset provides:

• —9,708 English Vision-QA rows (afri_aya_vision.jsonl)
• —19,416 Total Vision-QA rows including 13 African languages (afri_aya_vision_all.jsonl)
• —2,466 Unique African cultural images (extracted, bundled in images.zip, and embedded in Parquet)

Dataset Contents & Files

How to Use with Adaption

When uploading afri_aya_vision.jsonl to the Adaption platform:

Standard Vision Fine-Tuning

• —Prompt: prompt (or `prompt_vision` which includes `<image>`)
• —Completion: completion
• —Context Image: context_image (or `image_filename`)

Chat-Style VLM Fine-Tuning

• —Chat: chat (native `[{"role": "user", "content": "<image>\n..."}, {"role": "assistant", "content": "..."}]` list)
• —Context Image: context_image

How to Load in Python / Hugging Face datasets

from datasets import load_dataset

# Load directly from Hugging Face (images decoded automatically)
ds = load_dataset("gimmy256/afri-aya-vision", split="train")

print(ds[0])
# Image object: ds[0]['image']
# Question:     ds[0]['prompt']
# Answer:       ds[0]['completion']

Languages Covered

• —English (en)
• —Ganda / Luganda (lg)
• —Swahili (sw)
• —Yoruba (yo)
• —Zulu (zu)
• —Amharic (am)
• —Hausa (ha)
• —Igbo (ig)
• —Shona (sn)
• —Kinyarwanda (rw)
• —Chichewa (ny)
• —Sesotho (st)
• —Xhosa (xh)

Citation & Attribution

This dataset builds upon the work of the Cohere For AI Regional Africa Community:

@dataset{afri_aya_2024,
  author    = {Cohere Labs Regional Africa Community},
  title     = {Afri-Aya: Giving Sight to African LLMs},
  year      = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/CohereLabsCommunity/afri-aya}}
}